claude-real-video: Local video processing for any LLM

The utility claude-real-video (crv) prepares video for any model without cloud uploads. Using a yt-dlp link or a local file, it extracts frames and transcribes audio, outputting a folder containing frames, transcript.txt, and MANIFEST.txt.
Instead of fixed-interval sampling like Gemini’s default 1 frame per second, crv detects scene changes and removes duplicates using a sliding window. A 10-minute static slide, which would normally produce 600 near-identical frames, compresses to one, while fast cuts are captured frame-by-frame.
Key flags include --scene 0.30, --fps-floor 1.0, --max-frames 150, and --dedup-window 4. Transcription uses existing .srt or .vtt subtitles, or defaults to Whisper. Installation requires pip install claude-real-video, with [whisper] for audio and ffmpeg. It supports Python 3.10+ on macOS, Windows, and Linux under an MIT license.
Up next
tunnel-mcp enables direct communication between Claude agents
tunnel-mcp lets two Claude agents talk directly, skipping the human middleman. It uses a temporary cloudflared tunnel for the connection and NaCl secretbox for end-to-end encryption. The access link is single-use and expires in 10 minutes, keeping the session locked to two agents.
Related stories
- Tokendoctor cut its author’s Claude Code cost 38%
- Chisle’s author says it cuts oversized tool output by 48%
- Claude Desktop on Linux freezes over a 27-byte reply
- Claude Code Ultra week measured at 17B Opus cache reads
- memsaga gives Claude Code’s memory files a git history
- Gemini Omni 1.2 Flash drops 720p video to $0.07 a second
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

